| Authors: | B. Schneider, J. Dambre, P. Bienstman | | Title: | Reduction of analogue bandwidth requirement in photonic on-chip reservoir computing | | Format: | International Conference Proceedings | | Publication date: | 11/2012 | | Journal/Conference/Book: | Proceedings of the 2012 Annual Symposium of the IEEE Photonics Society Benelux Chapter
| | Editor/Publisher: | M. Wuilpart, C. Caucheteur and M. Mura, | | Volume(Issue): | p.227-230 | | Location: | Mons, Belgium | | Citations: | Look up on Google Scholar
| | Download: |
 (527KB) |
Abstract
Recurrent neural networks are brain-inspired dynamical systems that can compute cognitive tasks much more efficiently than traditional computing schemes. We study a special class of recurrent networks called Reservoir Computers (RC). This work demonstrates by means of simulations how an optical RC on chip can be made more tractable in terms of signal requirements. In fact, the short delay lines on a chip require a large analog bandwidth in order to achieve efficient computing. Using the example of the speech recognition task, we show that a fast digital masking operation together with additional feedback leads to a considerable decrease of the analog bandwidth requirements. Related Research Topics
Related Projects
|
|
